Why Your Coding Agents Need Program Design, Not Just Prompts

Dexter argues that coding agents solve problems but cannot produce maintainable code without human oversight. The key insight is that most teams skip "program design" – upfront decisions on architecture and call stacks before agents write code. This is a concrete signal: quality bottlenecks shift from coding to design.

Benchmarks mislead because they don't penalize sloppy code. Without a fast oracle for maintainability, reinforcement learning rewards passing tests, not clean architecture. Result: agents produce spaghetti code that becomes unmanageable over time. Hype around high SWE-bench scores should be met with skepticism.

The proposed solution is a four-stage workflow: product clarity, system architecture, program design, vertical slices. This requires more upfront human effort but reduces rework downstream. Tradeoff: more planning time vs less debugging later. Teams with product-market fit need this; pre-PMF startups can afford to move faster.

The "light software factory" where humans stop reading code leads to eventual disaster – a bug that agents can't fix and humans can't untangle. Counterintuitive: removing yourself from the loop increases long-term risk. Engineering leaders should invest in workflows that keep humans engaged in design and logic, not just code review.